Liquid AI launches LFM3 family of six open-weight 2b-class models - unite.ai. Liquid AI announced LFM3 on October 27, 2026, a new generation of open-weight Liquid Foundation Models that brings text, vision, audio, omni, and decision models onto one shared backbone. The first release, detailed in a post on the Hugging Face blog, covers six models in the 2B parameter range. Six models on one shared backbone. Every LFM3 model branches from the same LFM3-1.7B-Base. LFM3-1.7B takes text in and returns text. LFM3-VL-2B accepts text, images, and video and returns text, while LFM3-Audio-2B takes text or audio and responds with text or speech. LFM3-Omni-2B, which Liquid AI describes as its first omni model, accepts text, images, video, and audio and produces text and speech. Two decision models, d1-2B and d1-omni-2B, are specialized from the vision and omni models and return calibrated answers rather than generated text. According to the LFM3-1.7B model card, the text model has 1.73B parameters, a 256K-token context length, a 128,000-token vocabulary, and support for 16 languages, and it ships under the LFM Open License v1.0. The base and post-trained models are open-weight and available on Hugging Face, the Liquid AI Playground, and the Liquid AI API. Liquid AI's release post names day-0 launch partners Ambarella, AMD, Arm, Intel via OpenVINO, NVIDIA, OpenRouter, Qualcomm via Qualcomm AI Hub and GenieX, and Vercel, and says Mercedes-Benz and Shopify partnered on pre-release testing and evaluation. Tri-Hybrid architecture and training. LFM3-1.7B-Base has 24 layers: 13 gated short convolutions, five sliding-window attention layers with a 256-token window, and six global attention layers. The convolutions use a kernel size of 7, up from 3 in LFM2. Three further changes define the backbone: head-wise gated attention, in which a sigmoid gate scales each attention head's output and can suppress activation outliers that hurt quantization; partial RoPE, under which global attention layers rotate only the highest-frequency half of each head while sliding-window layers keep full RoPE; and a sandwich norm that normalizes each block's input and output. The design came from Liquid AI's hardware-in-the-loop architecture search, expanded for LFM3 to 14 device and backend combinations, including phones (Samsung Galaxy S26 Ultra and OnePlus 15), a Raspberry Pi 5, laptops (Apple M4 Max and AMD Ryzen AI), Arm and x86 cloud CPUs, GPUs, and the Qualcomm Hexagon NPU. LFM3-1.7B-Base was pre-trained on 22T tokens, with the final 3.5T tokens forming a mid-training stage on curated data with a larger share of agentic traces and tool calls that also extended the context window to 256K tokens. Post-training ran in four stages: supervised fine-tuning on 1.75T tokens, mixed reinforcement learning, multi-domain on-policy distillation from six specialist reinforcement-learning teachers, and agentic reinforcement learning inside real agent harnesses. Liquid AI said the new architecture lets LFM3-1.7B outperform its predecessor LFM2.5-2.6B with 35% fewer parameters, and that per parameter it matches LFM2-1.2B on short prompts while running about 25% faster with 8K tokens in context. At 4-bit precision, the company said, the text model takes about 1 GB of memory. Company-Reported performance and specialization. Liquid AI reports that LFM3-1.7B scores 38.42 on WebArena-Verified against 3.02 for LFM2.5-2.6B, with the largest gains on agentic tasks, though the model card states the model is not recommended for agentic coding or complex multi-step reasoning because of its small size. Liquid AI said LFM3-VL-2B is the first LFM vision model to accept video input, and reports that it raises the average across 62 vision benchmarks from 57.6 to 71.4 compared with LFM2.5-VL-3B while using 1.9B parameters against 3.1B, putting it on par with Gemma 4 12B, a model more than six times larger. A new in-house 164M-parameter vision encoder uses 3x less compute than the SigLIP2 encoder in previous LFM VL models and matches the accuracy of Qwen3.5-2B's vision encoder with 2.2x fewer FLOPs, according to the company's measurements. LFM3-Audio-2B supports 18 languages, and Liquid AI reports speech quality of 3.83 on UTMOSv2 against 3.22 for LFM2.5-Audio-1.5B, a fully causal audio encoder with time to first token as low as 50 ms, and a LiquidCodec tokenizer that generates speech at about 87x real time on a single CPU thread on an Apple M4 Max. The company said the 2.2B-parameter LFM3-Omni-2B beats models up to 10.7B parameters, including Qwen2.5-Omni-7B, on the Daily-Omni, WorldSense, and AV-SpeakerBench audio-visual understanding benchmarks. The decision models answer a list of questions in a single forward pass, each with a calibrated probability, and add typed answers that point at objects, read values from documents, estimate numbers with a range, and pick every option that applies. Liquid AI said d1-2B replaces d1-3B with a third fewer parameters and scores 75.5 against 64.0 on ImajevBench. Every LFM3 model can be fine-tuned on a single GPU, and Liquid AI published notebooks covering supervised fine-tuning, DPO, GRPO, and continued pre-training for the text, vision, audio, and omni models. In one company example, LFM3-VL-2B was fine-tuned on about 3,000 public wind-turbine blade photos for a few hours on one A100, raising defect classification accuracy to 67% and detection F1 from 0.003 to 0.11 on held-out photos. Liquid AI also adapted LFM3-VL-2B into three robotics models for orchestration, navigation, and manipulation. The company reports that the manipulation vision-language-action model, trained across 15 embodiments, achieved 96.2% average success across the four LIBERO suites and 45.7% on RoboCasa-GR1 tabletop tasks while generating action chunks in under 50 ms on a Jetson AGX Orin with TensorRT FP16. On throughput, Liquid AI said LFM3-1.7B can produce 3.6 billion tokens per day on a single NVIDIA B300 at 1,024 concurrent requests. The models ship with native support for llama.cpp, MLX, and vLLM across Apple, AMD, NVIDIA, and Qualcomm hardware, with checkpoint variants including BF16, FP8, NVFP4, GGUF, MLX, and ONNX.

Real-Time analytics AI gets a sub-50 ms push. October 22, 2026 28.4x is the biggest latency gap in Liquid AI's launch of RIPL, announced on October 22, 2026: in the company's own tests, one 1.2B model delivered 155.8 ms p95 versus 4,431 ms p95 for Qwen3.5-0.8B on vLLM. For enterprise teams tracking real-time analytics AI, that matters because it suggests generative models may be moving into request paths that were previously reserved for classical systems. According to MarkTechPost's coverage of the launch, the headline claim is sub-50 ms end-to-end responses for some live production workloads. That does not mean every request suddenly becomes fast enough for payments, search or fraud checks. It does mean the latency conversation is changing. If vendor-reported numbers hold up under buyer-side testing, the debate in 2027 may shift from can generative models fit live decision systems to which workloads justify the implementation and operating cost. Liquid AI launches RIPL for live request paths. Liquid AI's announcement is narrowly focused but strategically important. RIPL, short for Real-time Inference Pipeline, is designed to serve Liquid Foundation Models inside a live request path rather than in batch or long-running agent workflows. The company says some deployments can return complete responses end-to-end in under 50 ms while traffic continues at peak load. The hardware target is also clear. Liquid positions RIPL for enterprise deployments on NVIDIA H100, H200, B200 and RTX PRO 6000 GPUs, with published benchmarks run on a single B200. That matters because latency claims are often difficult to compare across serving stacks unless the hardware profile is explicit. NVIDIA's Blackwell platform materials give some context for why vendors are now pushing harder on production inference economics, not just model training. The practical message for buyers is simple: the company is not selling general-purpose serving flexibility first. It is selling low-latency behavior for specific production paths where every extra 50 to 100 milliseconds can affect revenue, approval rates or user abandonment. Why milliseconds matter more than model size in production. In enterprise systems, the key question is rarely whether a model is smart enough in the abstract. It is whether a decision can arrive inside a hard budget. Search reranking, fraud checks, payment authorization and adtech query rewriting do not get the luxury of multi-second waits. That is why classical models still dominate many of these systems today. Liquid AI's framing is credible on that point. Many open serving stacks, including vLLM, have been optimized around throughput, batching and broader model support. Those are useful goals, but they are not the same as serving a model in the critical path of a customer interaction. Shopify's mention is notable because it grounds the story in production relevance. Liquid says Shopify is deploying LFMs across latency-sensitive commerce workloads, including search enhancement at sub-20 ms inference. Shopify has separately documented how search quality and performance matter in its engineering ecosystem, and broader retail research such as Google's findings on mobile site speed has long shown that delays compound quickly. For teams evaluating architecture choices, this makes AI-powered data analytics dashboards a useful adjacent implementation model: the technical challenge is not just model quality, but how analytics, alerts and production decisions stay responsive under live demand. What the benchmark numbers actually show. The strongest part of the announcement is the specificity of the benchmark table. The weakest part is that the numbers are self-reported. Both things can be true. Here is the headline data Liquid published for p95 end-to-end latency on one B200 under open-loop Poisson arrivals: | Workload | RIPL latency | vLLM baseline | Reported gap | | 1B class, 32 in | 32 out | 34.6 ms | 498 ms | 14.4x | | 1B class, 256 in | 128 out | 155.8 ms | 4,431 ms | 28.4x | | 2B class, 32 in | 32 out | 76.4 ms | 586.8 ms | 7.7x | | 2B class, 256 in | 128 out | 449.1 ms | 2,499 ms | 5.6x | Three numbers stand out. * 34.6 ms p95 at 1,000 requests per second is the most commercially important result, because it falls into a range that starts to look viable for adtech and search edge cases. * 155.8 ms p95 is still above the launch's sub-50 ms headline, but it remains relevant for workflows like risk scoring or assisted decisioning where 150 ms is acceptable and 4 seconds is not. * The drop from a 28.4x advantage to 5.6x on longer 2B-class workloads is the reminder that prompt length and output size still govern economics. This is where enterprise buyers should stay balanced. According to Liquid AI's source post, as relayed by MarkTechPost, RIPL appears to beat vLLM even when running the same LFM2.5-350M model, which suggests the engine itself contributes to the gain. But the comparison is still vendor-framed, with different model families, different design choices and a closed enterprise engine on one side. How RIPL differs from a normal serving stack. The notable claim is not only that LFM2.5 models are fast. It is that RIPL optimizes the entire request path from arrival to completed response. That is an architectural claim, not just a model benchmark. Liquid describes LFM2 as a hybrid family mixing double-gated convolution blocks with grouped-query attention blocks. On paper, that matters because lower-latency inference is often the result of cumulative engineering decisions: token generation path, memory movement, scheduler behavior, queue management and model architecture. The MLPerf Inference benchmarks have repeatedly shown that system-level performance is rarely explained by model weights alone. For enterprise AI integrations, this is the more interesting signal. Many teams still think of generative inference as a model-selection problem. In practice, the production outcome depends just as much on serving architecture, concurrency assumptions and failure handling. A good model in the wrong stack can still miss the business SLA. Which enterprise workflows fit RIPL best. The six workload examples in the launch tell a clearer story than the headline alone. Liquid maps RIPL to adtech query rewriting, search intent classification, search reranking, payments risk scoring, credit decisioning and image alt text generation. Those are not toy demos. They are revenue-adjacent decisions with narrow budgets. The fit is strongest where two conditions hold: the request is short enough to keep latency controlled, and the business benefit of generative reasoning is higher than what a classical model already delivers. That is especially relevant in AI for fintech and AI fraud detection, where richer context can improve decision quality but only if the model returns quickly enough to preserve throughput. A useful way to read the list is by latency tolerance: * 20 to 30 ms budgets: search intent classification and reranking look plausible only for small, sharply optimized workloads. * 50 ms budgets: adtech rewriting and payments risk scoring are the real proving ground for real-time analytics AI. * 100 to 150 ms budgets: image alt text and credit decisioning broaden the commercial case, even without the sub-50 ms headline. That breakdown also explains why the launch matters more to e-commerce, payments and adtech than to teams running slower internal copilots. If the traffic path is customer-facing and high-volume, shaving hundreds of milliseconds is operationally meaningful. If the workflow is already asynchronous, the advantage narrows. What this launch signals for enterprise AI teams. The trend line is becoming harder to ignore. In 2025, many enterprises treated generative models as assistants outside the transaction path. By late 2026, vendors are now presenting them as candidates for production decisions that happen in milliseconds. RIPL does not settle the question, but it does move the line. The operational lesson is straightforward. Buyers should test for p95 and p99 latency under continuous load, not average latency on clean demos. They should also compare short and long prompts separately, because Liquid's own figures show how fast the economics change as token counts rise. And they should ask whether closed serving software creates dependencies that outweigh the performance gain. For real-time analytics AI, the bigger trend is not that one engine posted a fast number. It is that implementation teams may soon have fewer reasons to default to classical models for every live decision path. The remaining work is disciplined validation, not vendor applause. Co-Founder & CEO, encorp.ai CEO and Founder of Encorp.io with expertise in AI and business transformation

Liquid AI's d1-ultra takes the top spot on Artificial Analysis Decision Index. The new decision model scores 65 by refusing to make things up, even if it gets fewer answers right an hour ago Sponsored: Vera - AI-powered prediction market intelligence, built for serious analysts Explore Vera Liquid AI has a new leaderboard champion, and it won by knowing what it doesn't know. The company's d1-ultra model now sits first on the Artificial Analysis Decision Index with a score of 65. It pairs a high non-hallucination rate with lower accuracy, which is a trade-off some businesses will find very appealing. What d1-ultra actually does. Liquid AI announced d1-ultra on October 15, 2026. It is a multimodal decision model, meaning it handles both text and images. Instead of generating free-form text, d1-ultra looks at a predefined set of answer options and assigns each one a calibrated probability. Picture a multiple-choice exam where the student also writes how confident they are next to every bubble. Calibration is the key word. A well-calibrated model that says it's 80% sure should be right about 80% of the time, not 40%. That matters for the use cases Liquid AI is targeting. The company points to tasks like visual defect detection and insurance claim verification, plus policy checks. In those settings, a model that flags uncertainty is useful. Confident cases get automated, and shaky ones get handed off to a human operator. That is the workflow Liquid AI is pitching. Under the hood, the model processes text and images together in a single forward pass, then outputs its odds without generating free-form output tokens. AI, tech, and the markets they move - in one daily briefing. Daily. Free. Join 34,000+ readers across crypto, finance, and policy. The numbers behind the ranking. d1-ultra's 65 on the Decision Index puts it 8 points ahead of the prior leader, Jev 1.14, which scored 57. An 8-point gap on a freshly launched index is a comfortable lead. It is also a lead on a leaderboard that has existed for roughly two days, since the Decision Index began public operations on October 13, 2026. The index evaluates models across four distinct assessments: * Knowledge and hallucination: does the model know facts, and does it avoid inventing them * Reasoning: can it work through multi-step problems * Tool usage: can it operate within systems that call external functions * Safety classification: can it correctly sort content by risk Separately, Liquid AI ran its own 112-evaluation suite. According to the company, d1-ultra outperformed both the earlier d1 model and Jev 1.14 across all eight evaluation categories in that suite. Liquid AI also says d1-ultra posted the lowest calibration error across those 112 evaluations. Pricing, languages and availability. d1-ultra is available through the Liquid API at $0.042 per million input tokens. The model supports 44 languages. Integrations with Vercel and OpenRouter are expected, though Liquid AI has said those will be announced later. For now, the Liquid API is the front door. What this means. The most important caveat is verification. Independent verification of Liquid AI's internal evaluations has not been established, which leaves some doubt about how firm the absolute rankings on the new index really are. Still, the design choice behind d1-ultra is worth taking seriously. In regulated, high-stakes workflows, a wrong answer delivered confidently can cost far more than no answer at all. An insurer doesn't need a model that approves every claim with swagger. It needs one that clears the obvious cases and escalates the weird ones. A model trading some accuracy for fewer hallucinations may be the better fit there, even if it looks weaker on a pure accuracy chart. Disclosure: This article was edited by Diego Almada Lopez. For more information on how Cryptobriefing create and review content, see its Editorial Policy.